Performance Analysis Results

This section lists performance analysis experiments conducted focusing on stream query processing and event ingestion with persistence.

1 Summary
2 Stream query processing
3 Ingesting events with persistence

Summary

The following table shows a complete summary of the content presented in this section.

Tested Item	Data Store	Query Type	Amount of events processed	Average Throughput (events per second)	Average Latency (milliseconds)	Message Payload Size (bytes)

Tested Item	Data Store	Query Type	Amount of events processed	Average Throughput (events per second)	Average Latency (milliseconds)	Message Payload Size (bytes)
Simple Pass-through	None	None	30 million	900K	0.9	12
Filter	None	Filter out all the events	30 million	900K	1.5	12
Window-small (1 second)	None	Sliding time window	30 million	100K	48	12
Window - Large (1 minute)	None	Sliding time window	30 million	100K	130	12
Patterns	None	Temporal event sequence patterns	1250 million	500K	550	76
Event Ingestion with Persistence	Oracle Event Store	Insert	252 million	70K	42	40
	Oracle Event Store	Update	75 million	20K	12	40
	MS SQL Event Store	Insert	198 million	55K	44.2	40
	MS SQL Event Store	Update	3.6 million	1K	4.6	40
	MySQL Event Store	Insert	12.2 million	3.4K	2.14	40
	MySQL Event Store	Update	3 million	500	0.5	40

Notes

Event ingestion with persistence tests were conducted using the default Amazon RDS configurations.
All the tests for event ingestion with persistence were conducted for 1 hour.
Performance results were aggregated in 5000K event windows.
The above table had an input rate of 1000K events per second during the first four tests.
All the tests were conducted using TCP transport.

1 Summary
2 Stream query processing
- 2.1 Scenario: Running multiple Siddhi queries
- 2.2 Scenario: Running Siddhi Pattern query on debs-2013-grand-challenge-soccer-monitoring dataset
3 Ingesting events with persistence

Stream query processing

Scenario: Running multiple Siddhi queries

Infrastructure used

The experiments were carried out in two c4.2xlarge (8 vCPU, 16GB RAM, EBS storage with 1000 Mbps max dedicated bandwidth) Amazon EC2 instances.
- Linux kernel 4.44, java version "1.8.0_131", JVM flags -Xmx4g -Xms2g
One node operated as a client.
Another node operated as a Stream Processor node.
Experiments were carried out using TCP as the transport.

Query Type	Sample Query	Amount of Events Processed	Average Throughput (events per second)	Latency (ms)

Query Type	Sample Query	Amount of Events Processed	Average Throughput (events per second)	Latency (ms)
Simple Passthrough	`@App:name("TCP_Benchmark")` `@source(type = 'tcp', context='inputStream',@map(type='binary'))` `define stream inputStream (iijtimestamp long,value float);` `from inputStream` `select iijtimestamp,value` `insert into tempStream;`	30 million	900K	0.9
Filter	`@App:name("TCP_Benchmark")` `@source(type = 'tcp', context='inputStream',@map(type= 'binary'))` `define stream inputStream (iijtimestamp long,value float);` `from inputStream[value<=1]` `select iijtimestamp,value` `insert into tempStream;`	30 million	900K	1.5
	`@App:name("TCP_Benchmark")` `@source(type = 'tcp', context='inputStream',@map(type= 'binary'))` `define stream inputStream (iijtimestamp long,value float);` `from inputStream[value<=0.5]` `select iijtimestamp,value` `insert into tempStream;`	30 million	450K	1.1
	`@App:name("TCP_Benchmark")` `@source(type = 'tcp', context='inputStream',@map(type= 'binary'))` `define stream inputStream (iijtimestamp long,value float);` `from inputStream[value<=0.25]` `select iijtimestamp,value` `insert into tempStream;`	30 million	226K	0.6
Window	`@App:name("TCP_Benchmark")` `@source(type='tcp',context='inputStream',@map(type='binary'))` `define stream inputStream (iijtimestamp long,value float);` `from inputStream#window.time(1 sec)` `select iijtimestamp,value` `insert into tempStream;`	30 million	100K	48
Window	`@App:name("TCP_Benchmark")` `@source(type='tcp',context='inputStream',@map(type='binary'))` `define stream inputStream (iijtimestamp long,value float);` `from inputStream#window.time(1 min)` `select iijtimestamp,value` `insert into tempStream;`	30 million	100K	130

Scenario: Running Siddhi Pattern query on debs-2013-grand-challenge-soccer-monitoring dataset

Infrastructure used

The experiments were carried out in two c4.2xlarge (8 vCPU, 16GB RAM, EBS storage with 1000 Mbps max dedicated bandwidth) Amazon EC2 instances.
- Linux kernel 4.44, java version "1.8.0_131", JVM flags : -Xmx4g -Xms2g
One node operated as a client.
Another node operated as a Stream Processor node.
Experiments were carried out using TCP as the transport.

Dataset

The data used in 2013 DEBS Grand Challenge is collected by the Real-Time Locating System deployed on a football field of the Nuremberg Stadium in Germany. Data originates from sensors located near the players’ shoes (1 sensor per leg) and in the ball (1 sensor). The goalkeeper is equipped with two additional sensors in each hand. The sensors in the players’ shoes and hands produce data at a frequency of 200Hz, while the sensor in the ball produces data at a frequency of 2000Hz. The total data rate reaches roughly 15 position events per second. Every position event describes the position of a given sensor in a three-dimensional coordinate system. The center of the playing field is at coordinate (0, 0, 0) for the dimensions of the playing field and the coordinates of the kickoff.

For more details about the dataset, see DEBS 2013 Grand Challenge: Soccer monitoring.

Pattern used

We created patterns for goal scoring scenario. In this scenario, the pattern matching query involves two events referrred to as e1 and e2 that should occur one after the other with some preconditions being satisfied. These preconditions include the position (denoted by x, y, and z) and the acceleration of the ball (denoted by a_abs). The numerical constants (such as 29880, 22560, etc.,) in the sample pattern matching query with which the values for x, y, and z are compared correspond to the boundary points of the goal reg.

Sample Siddhi App

@App:name("TCP_Benchmark")

@source(type = 'tcp', context='inputStream',@map(type='binary'))
define stream innerStream(iij_timestamp long,sid int, eventtt long, x double, y, double, z int, v_abs double, a_abs int, vx int, vy int, vz int, ax int,ay int, az int);

from  e1=innerStream[(x>29880 or x<22560) and y>-33968 and y<33965 and (sid==4 or sid ==12 or sid==10 or sid==8)]
    -> e2=innerStream[(x<=29898 and x>22579) and y<=-33968 and z<2440 and a_abs>=55000 and (sid==4 or sid ==12 or sid==10 or sid==8)]

select
 e2.sid as sid, e2.eventtt as eventtt, e2.x as x,e2.y as y, e2.z as z,e2.v_abs as v_abs,e2.a_abs as a_abs, e2.vx as vx,e2.vy as vy, e2.vz as
 vz, e2.ax as ax,e2.ay as ay, e2.az as az, e1.iij_timestamp as iij_timestamp
insert into outputStream;

Summary Results

Throughput (events per second)	500,000
Latency (ms)	550

Ingesting events with persistence

All the event ingestion with persistence tests were conducted using the following deployment configuration.

Oracle event store

Infrastructure used

The following infrastructure was used in both scenarios:

c4.2xlarge (8 vCPU, 16GB RAM, EBS storage with 1000 Mbps max dedicated bandwidth) Amazon EC2 instances operated as the SP node and the TCP client.
- Linux kernel 4.44, java version "1.8.0_131", JVM flags : -Xmx4g -Xms2g
db.m4.2xlarge (8 vCPU, 32 GB RAM, EBS-optimized storage with 100 Mbps max dedicated bandwidth) Amazon RDS instance with Oracle operated as the database node.
Customized TCP client operated as the data publisher (TCP producer found in samples).
Experiments were carried out using Oracle 12g.

Scenario 1: Insert query - Persisting 252 million events of process monitoring events in Oracle

This test involved persisting process monitoring events of approximately 180 bytes each. The test injected 252 million events into WSO2 Stream Processor with a publishing TPS of 70,000 events per second during a time period of one hour.

Throughput Graph

Latency Graph

Summary Results

Throughput (events per second)	70,000
Latency (ms)	42

Scenario 2: Update Query - Updating 10 million events in Oracle Data store

The test injected 10 million events into the properly indexed Oracle Database. The test involved persisting process monitoring events of approximately 180 bytes each. 75 million update queries were performed. The publishing throughput was 20,000 events per second during a time period of one hour.

Throughput Graph

Latency Graph

Summary Results

Number Of Persisted Events	10 million
Throughput (events per second)	20,000
Latency (ms)	12

Microsoft SQL server event store

Infrastructure used

c4.2xlarge (8 vCPU, 16GB RAM, EBS storage with 1000 Mbps max dedicated bandwidth) Amazon EC2 instance operated as the SP node.
- Linux kernel 4.44, java version “1.8.0_131", JVM flags : -Xmx4g -Xms2g
db.m4.2xlarge (8 vCPU, 32 GB RAM, EBS-optimized storage with 100 Mbps max dedicated bandwidth) Amazon RDS instance with MS SQL Enterprise Edition 2016 operated as the database node
Customized TCP client operated as the data publisher (Sample TCP client found in samples).

Scenario 1: Insert Query - Persisting 198 million process monitoring events in MS SQL

This test involved persisting process monitoring events of approximately 180 bytes each. The test injected 198 million events into WSO2 Stream Processor with a publishing throughput of 55,000 events per second during a time period of one hour.

Throughput Graph

Latency Graph

Summary Results

Throughput (event per second)	55,000
Latency (ms)	44.2

Scenario 2: Update Query - Updating 10 million events in MS SQL data store

This test injected 10 million events into the properly indexed MS SQL Database. This test involved persisting process monitoring events of approximately 180 bytes each. 3.6 million update queries were performed with a publishing throughput of 1000 events per second during a time period of one hour.

Throughput Graph

Summary Results

Number Of Persisted Events	10 million
Throughput (events per second)	1000
Latency (ms)	4.6

MySQL event store

Infrastructure Used

c4.2xlarge (8 vCPU, 16GB RAM, EBS storage with 1000 Mbps max dedicated bandwidth) Amazon EC2 instances operated as the SP node and the TCP client.
- Linux kernel 4.44, java version "1.8.0_131", JVM flags : -Xmx4g -Xms2g
db.m4.2xlarge (8 vCPU, 32 GB RAM, EBS-optimized storage with 100 Mbps max dedicated bandwidth) Amazon RDS instance with MySQL Community Edition version 5.7 operated as the database node.
Customized TCP client operated as the data publisher (TCP producer found in samples).
Experiments were carried out using 5.7.19 MySQL Community Server.

Scenario 1: Insert Query - Persisting 12.2 million process monitoring events in MySQL

This test involved persisting process monitoring events of approximately 180 bytes each. The test injected 12.2 million events into WSO2 Stream Processor with a publishing throughput of 3400 events per second during a time period of one hour.

Throughput Graph

Latency Graph

Summary Results

Throughput (events per second)	3400
Latency (ms)	2.14

MySQL Upper Limit

After about 12.2 million events are published, a sudden drop can be observed in the receiver performance. This number can be considered as the upper limit of MySQL event store with default settings. In order to continue receiving events without a major performance degradation, data should be purged periodically from the event store before it reaches the upper limit.

Scenario 2: Update Query - Updating 100K events in MySQL data store

The test injected 100,000 events into the properly indexed My SQL Database. This test involved persisting process monitoring events of approximately 180 bytes each. Three million update queries were performed with a publishing throughput of 500 events per second during a time period of one hour.

Throughput Graph

Latency Graph

Summary Results

Number Of Persisted Events	100K
Throughput (events per second)	500
Latency (ms)	0.5

Conclusion

The performance test results indicate that the event persistence performance of WSO2 Stream Processor is characterized by the event store database performance.